def __init__(self, relext_model_name, models_foldes="../trainedmodels/"):
module_location = os.path.abspath(__file__)
module_location = os.path.dirname(module_location)
model_params = keras_models.model_params
max_sent_len = keras_models.model_params['max_sent_len']
self._embeddings, self._word2idx = embeddings.load(
keras_models.model_params['wordembeddings'])
print("Loaded embeddings:", self._embeddings.shape)
self._idx2word = {v: k for k, v in self._word2idx.items()}
self._model = getattr(keras_models, relext_model_name)(
model_params, np.zeros((len(self._word2idx), 50),
dtype='float32'), max_sent_len,
len(keras_models.property2idx))
self._model.load_weights(models_foldes + relext_model_name +
".kerasmodel")
with codecs.open(os.path.join(
module_location, "../../resources/properties-with-labels.txt"),
encoding='utf-8') as infile:
self._property2label = {
l.split("\t")[0]: l.split("\t")[1].strip()
for l in infile.readlines()
}
self._graphs_to_indices = keras_models.to_indices
if "Context" in relext_model_name:
self._graphs_to_indices = keras_models.to_indices_with_extracted_entities
elif "CNN" in relext_model_name:
self._graphs_to_indices = keras_models.to_indices_with_relative_positions
def __init__(self, relext_model_name, models_foldes="../trainedmodels/",
embeddings_location="glove/glove.6B.50d.txt", resource_folder="../resources/"):
with open(models_foldes + relext_model_name + ".property2idx") as f:
self._property2idx = ast.literal_eval(f.read())
module_location = os.path.abspath(__file__)
module_location = os.path.dirname(module_location)
with open(os.path.join(module_location, "../model_params.json")) as f:
model_params = json.load(f)
self._embeddings, self._word2idx = embeddings.load(embeddings_location)
print("Loaded embeddings:", self._embeddings.shape)
self._idx2word = {v: k for k, v in self._word2idx.items()}
self._model = keras_models.model_ContextWeighted(model_params,
np.zeros((len(self._word2idx), 50), dtype='float32'),
max_sent_len, len(self._property2idx))
self._model.load_weights(models_foldes + relext_model_name + ".kerasmodel")
with open(resource_folder + "properties-with-labels.txt") as infile:
self._property2label = {l.split("\t")[0] : l.split("\t")[1].strip() for l in infile.readlines()}
self._idx2property = {v: k for k, v in self._property2idx.items()}
self._graphs_to_indices = keras_models.to_indices_with_real_entities
if "CNN" in relext_model_name:
self._graphs_to_indices = keras_models.to_indices_with_relative_positions
def __init__(self, relext_model_name, models_folder="../trainedmodels/"):
"""
Initialize a new relation parser with the given model type. This class simplifies the loading of models and
encapsulates encoding sentences into the correct format for the given model.
:param relext_model_name: The name of the model type that should correspond to the correct model class and
the name of the model file
:param models_folder: location of pre-trained model files
"""
module_location = os.path.abspath(__file__)
module_location = os.path.dirname(module_location)
model_params = keras_models.model_params
max_sent_len = keras_models.model_params['max_sent_len']
self._embeddings, self._word2idx = embeddings.load(
keras_models.model_params['wordembeddings'])
print("Loaded embeddings:", self._embeddings.shape)
self._idx2word = {v: k for k, v in self._word2idx.items()}
self._model = getattr(keras_models, relext_model_name)(
model_params, np.zeros((len(self._word2idx), 50),
dtype='float32'), max_sent_len,
len(keras_models.property2idx))
self._model.load_weights(models_folder + relext_model_name +
".kerasmodel")
with codecs.open(os.path.join(
module_location, "../../resources/properties-with-labels.txt"),
encoding='utf-8') as infile:
self._property2label = {
l.split("\t")[0]: l.split("\t")[1].strip()
for l in infile.readlines()
}
self._graphs_to_indices = keras_models.to_indices
if "Context" in relext_model_name:
self._graphs_to_indices = keras_models.to_indices_with_extracted_entities
elif "CNN" in relext_model_name:
self._graphs_to_indices = keras_models.to_indices_with_relative_positions
)
parser.add_argument('--property_index')
parser.add_argument(
'-s',
action='store_true',
help="Use only a portion of the training and validation sets.")
args = parser.parse_args()
model_name = args.model_name
mode = args.mode
with open(args.model_params) as f:
model_params = json.load(f)
embeddings, word2idx = embeddings.load(args.word_embeddings)
print("Loaded embeddings:", embeddings.shape)
training_data, _ = io.load_relation_graphs_from_file(args.train_set,
load_vertices=True)
val_data, _ = io.load_relation_graphs_from_file(args.val_set,
load_vertices=True)
if args.s:
training_data = training_data[:len(training_data) // 3]
print("Training data size set to: {}".format(len(training_data)))
val_data = val_data[:len(val_data) // 3]
print("Validation data size set to: {}".format(len(val_data)))
if mode in ['test', 'train-plus-test']:
if __name__ == "__main__":
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('model_name')
parser.add_argument('mode', choices=['train', 'optimize', 'train-continue'])
parser.add_argument('train_set')
parser.add_argument('val_set')
parser.add_argument('--models_folder', default="../trainedmodels/")
args = parser.parse_args()
model_name = args.model_name
mode = args.mode
embedding_matrix, word2idx = embeddings.load(keras_models.model_params['wordembeddings'])
print("Loaded embeddings:", embedding_matrix.shape)
training_data, _ = io.load_relation_graphs_from_file(args.train_set, load_vertices=True)
val_data, _ = io.load_relation_graphs_from_file(args.val_set, load_vertices=True)
print("Training data size: {}".format(len(training_data)))
print("Validation data size: {}".format(len(val_data)))
max_sent_len = keras_models.model_params['max_sent_len']
print("Max sentence length set to: {}".format(max_sent_len))
to_one_hot = np_utils.to_categorical
graphs_to_indices = keras_models.to_indices
if "Context" in model_name:
to_one_hot = embeddings.timedistributed_to_one_hot
import os
import numpy as np
import tensorflow as tf
import random
import sys
import json
import _pickle as cPickle
from core import embeddings
embedding_matrix, word2idx = embeddings.load('./glove/glove.6B.50d.txt')
print("Loaded embeddings:", embedding_matrix.shape)
### preprocess all the wiki data and select desired data to construct data pairs
CORPUS_FOlDER = "./enwiki-20160501"
propertylist = [
'P20', 'P19', 'P551', 'P463', 'P108', 'P157', 'P69', 'P172', 'P140', 'P26',
'P40', 'P22', 'P25', 'P119', 'P66', 'P27', 'P101', 'P800', 'P166', 'P39',
'P102', 'P263', 'P184', 'P802', 'P53', 'P553', 'P1344', 'P1416', 'P103',
'P91', 'P237', 'P411', 'P412', 'P450', 'P97', 'P512', 'P1303', 'P1399',
'P1412', 'P1429', 'P451', 'P1038', 'P21', 'P734', 'P735', 'P570', 'P569',
'P1196', 'P106', 'P509'
]
relation_num = {
'P20': 50,
'P19': 1,
'P551': 2,
'P463': 3,
'P108': 4,
'P157': 5,
'P69': 6,
'P172': 7,
# parser.add_argument('val_set')
parser.add_argument('--models_folder', default="./trainedmodels/")
parser.add_argument('--earlystop', action='store_true')
parser.add_argument('--epoch', default=50, type=int)
parser.add_argument('--checkpoint', action='store_true')
parser.add_argument('--tensorboard', action='store_true')
parser.add_argument('--metadata', type=str)
parser.add_argument('--error_out_folder', default='./error_output/')
args = parser.parse_args()
model_name = args.model_name
mode = args.mode
error_out_folder = args.error_out_folder
embedding_matrix, word2idx = embeddings.load(args.word_embedding)
print("embedding_matrix: " + str(embedding_matrix.shape))
if args.exist:
relationMention_ph = args.train_set
train_data, val_data, test_data = io.load_relation_from_existing_sets(
relationMention_ph)
else:
relationMention_files = glob.glob(args.train_set)
train_data, val_data, test_data = io.load_relation_from_files(
relationMention_files, val_portion=0.1, test_portion=0.1)
print("Document number: {}".format(len(relationMention_files)))
if mode == 'create-data-set':
with open('./data/relationMention/train.relationMention.json',
'w') as f:
